LTX Video

LTX Video (LTXV) comes from Lightricks, the team behind the LTX-Video research line. It is built on a diffusion transformer (DiT) rather than the U-Net used by the SD/SDXL lineage, and it was the first DiT-based video model able to generate high-quality clips in real time — the original release produces 24 FPS video faster than it can be watched. Like other transformer models it reads long, natural-language descriptions through a T5 text encoder, and it handles both text-to-video and image-to-video from a single starting frame.

The family has grown from a fast, lightweight base into a full audio-visual foundation model. The 2B model (0.9.x) is the light, near-real-time option; LTX Video 2 (19B) is a joint audio-video model that generates synchronized video and audio together within one network; and LTX Video 2.3 is a significant update to LTX-2 with improved audio and visual quality and stronger prompt adherence. All are open-weight and built for practical local execution, and the ecosystem has already spawned community fine-tunes — such as the uncensored Sulphur 2, built directly on the LTX 2.3 pipeline.

Choose LTXV when iteration speed matters most: its diffusion-transformer design turns prompts and starting images into coherent clips far faster than heavier video models, and newer versions add synchronized audio in the same pass. For the deepest LoRA library and the most detailed motion control, the Wan ecosystem still leads; Hunyuan Video leans toward cinematic realism, and Kling is a closed cloud API. Because even the larger LTXV models are engineered for efficiency, they generally cost less Buzz per second of video than the heaviest open alternatives.